The Navigator’s Journal: Observations from leaders navigating AI, strategy, and organizational change.
Why Trusting AI Agents Starts with Deciding Who Gives the Order
Here’s a question worth asking before your next AI project. When the tool stops answering and starts acting, who’s holding the helm?
Over the last few years, most of us have used AI the way a navigator at sea uses a chart. You consult it, weigh it, and then someone makes the call. You ask, it answers, you decide. That arrangement is changing.
The newest tools, called AI agents, don’t just recommend a course. They take the wheel. They fill in forms, move data between systems, and send messages. Platforms such as Claude can literally integrate with Chrome and click through buttons on your behalf. Agents are also expected to remain one of the biggest AI themes this October.
If you are running a business, that raises a practical question. How much of the wheel do you hand over, and on what terms? We’ve noticed that most leaders ask which agent to buy. We think the better question comes first. What are we actually trying to solve, and who stays accountable while it’s being solved?
The trust gap is really a seamanship gap
What does the research say? A Workato and Harvard Business Review survey of more than 600 tech leaders found that only 6% fully trust agents with essential end-to-end business processes. Another 43% trust agents only with routine operational tasks, and 8% don’t trust them with business operations at all. That survey was published in late 2025, and the picture looks similar in 2026.
Security sits at the center of the hesitation. AvePoint’s 2026 survey of 750 global IT leaders ranks security and privacy as the top concerns for both generative AI and AI agents. It also found organizations are delaying deployments mainly over unresolved data security and data management concerns.
Smaller businesses feel it too. A Simply Business outlook reports 62% of small business owners have already started using AI, yet 34% cite data security as a hesitation, specifically where sensitive information goes once it’s typed into a prompt.
At SpecNav, we don’t read that as fear of technology. We read it as seamanship. No competent crew sends an untested hand to the helm in a busy channel. The caution leaders are showing is the same instinct that keeps a ship off the rocks. The risk isn’t that leaders are too careful. The risk is that agents arrive in the organization one convenient feature at a time, with nobody having decided how much authority they’ve been given.
Float before you move
This is where SpecNav’s philosophy earns its keep. Float. Move. Sustain. This starts with stabilizing the organization before anything changes. With agents, that means settling a few things before the first one touches a live system.
The relationship between a Captain and Navigator is a useful relationship to compare to. The navigator reads the water, plots the course, and calls out hazards early. The captain gives the order. An agent can be a very capable navigator. But it should not quietly become the captain. Your leadership, and your accountability to customers, staff, and regulators, can’t be delegated to software, however fluent it sounds.
AI is the tool. Leadership is the craft. The craft here is deciding, deliberately, how much authority each task earns. To do that, the process, and the challenges or blockers within it must be fully understood.
Three rungs of autonomy
We find it helps to think in rungs of a ladder and not a single on/off switch. This is a framework to adapt, not a statistic. Your business may need more rungs or fewer.
Rung one: Chart it. The agent gathers information, drafts, summarizes, and recommends. A person reviews everything and acts on it. The answers stay in the AI environment and must be consciously exported so that nothing leaves the building without human hands on it. Most organizations should start here, and many tasks should stay here. Automatically scaling to fully autonomous processes significantly increases the business’ risk surface.
Rung two: Plot it, hold for the order. The agent prepares the action, such as a drafted reply, a filled form, or a proposed schedule change, in the actual system and then waits. A named person approves before anything is sent, posted, paid, or changed. The agent does the legwork and the human keeps the decision and the responsibility.
Rung three: Standing orders within limits. The agent acts on its own inside a tightly defined boundary, and every action is logged so a person can review it. That means it reviews, drafts, confirms and sends. A ship runs this way on a routine watch. The captain’s standing orders define what the watch can do alone and when to wake the captain. For AI, that is the escalation pathway.
One rule keeps the ladder honest. Nothing moves up a rung until it has earned it on the one below. If an agent hasn’t produced reliable drafts at rung one, it hasn’t earned the right to act at rung three. And some work probably never belongs on the top rung. Anything irreversible, anything that speaks for you to a customer in a sensitive moment, and anything touching money or regulated information deserves a person at the helm. Compliance questions in particular are ones to take to your legal or compliance advisors, not to settle by default based on loose assumptions and simply because “AI can”. The question is, “should AI”?
Four questions to answer at the helm
Before an agent takes any task, we’d suggest answering these four questions in plain language and in writing. They draw on the BATHERS AI ethical heck from our recent white paper, Problem First, Tool Second.
- Who owns the outcome? This is Accountability and can only really be achieved when there is a Human in the Loop. If the agent gets it wrong, a named person owns the fix. “The tool did it” is not an answer a customer will accept.
- What can it see? This is Safeguards, relating to Security and Privacy. Agents work with whatever access you give them. Give an agent the keys to everything and a single mistake, or a single compromise, travels a long way. Your risk surface has become exponentially bigger. Scope access to the task, the way you’d scope a new hire’s access in their first week. What does the Agent need to see to be able to do the basics of its job only?
- What can it do, and can it be undone? This is human oversight. Write down the actions it may take and the ones it may not. These are the guardrails. The Captain’s Standing Orders. Prioritize actions you can reverse. If a step can’t be undone, it belongs at rung two or lower in order to reduce the potential risk.
- How will we know what happened? This is transparency. You should be able to see what the agent did, in what order, and why. If you can’t reconstruct a day’s work from the log, the agent is running dark, and nobody should be comfortable with that on any watch.
None of these questions requires a technical team. They require a leader who has stopped to ask them. For owner-operators running the whole business without an IT department, that’s good news. Control is much more easily managed especially where a few, key agents are in place. But scaled in a larger business or organization, with multiple departments and multiple outputs. The risk becomes much greater as does the need for governance. This is when a culture of discipline matters more than the budget.
Where to start on Monday
You don’t immediately need an agent strategy for the whole company. That is part of the roadmap and maturity model. It’s something to aim toward in the future. On Monday, you need one well-chosen first step.
- Identify a recognized business challenge. We all have them. The part of a process that doesn’t quite work or the manual, administrative task that takes more time than it should.
- Pick one routine, low-risk task. Based on the challenge, choose something involved that is repetitive that costs you real time and where a mistake is cheap to catch, such as sorting incoming inquiries or preparing a first draft of a weekly report.
- Write the standing order. In a few plain sentences, say what the agent may do, what it may touch, and when it must stop and ask.
- Run it at rung one for a couple of weeks. Compare what it produces with what a person would have done. Note where it surprised you, in either direction.
- Review the record, then decide. Only then consider rung two. Keep a person in the loop on anything that reaches a customer.
You may find the first agent task is the least interesting part. The bigger value often comes from being forced to write down who owns what, who can see what, and what happens when something goes wrong. Organizations that haven’t done that work tend to find the gap the hard way.
Steady hands, clear orders
In Chart 005, we looked at the gap between how fast technology moves and how fast people adapt. Agents widen that gap, because they’re the first tools that act without waiting to be asked. Charts 001 to 005 point to the same conclusion – leadership readiness comes before tool selection.
So the question for your next planning meeting isn’t whether to use agents. It’s this. Where, specifically, would you be comfortable letting one hold the wheel, and who gives the order that puts it there?
We’re curious. What conversations are happening in your organization about AI agents right now? 🧭
AI-assisted: This article was researched and drafted with AI assistance and reviewed by a human before publication.
